RoboInter/RoboInterTools/tools/sam.py at main · InternRobotics/RoboInter · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
import cv2
import torch
import numpy as np
import matplotlib.pyplot as plt
from sam2.build_sam import build_sam2_video_predictor


class Sam:
    def __init__(
        self,
        sam2_checkpoint,
        model_cfg,
        threshold=0.0,
        save_visualization=False,
        device="cpu",
    ):
        self.sam2_checkpoint = sam2_checkpoint
        self.model_cfg = model_cfg
        self.threshold = threshold
        self.save_visualization = save_visualization
        self.device = device
        self.predictor = build_sam2_video_predictor(
            self.model_cfg, self.sam2_checkpoint, device
        )

    def reset_predictor(self):
        self.predictor = build_sam2_video_predictor(
            self.model_cfg, self.sam2_checkpoint, self.device
        )
        self.inference_state = None

    def get_mask_on_image(
        self, masks_list, video, obj_id=None, random_color=False, save_path=None
    ):
        if self.save_visualization:
            assert (
                save_path is not None
            ), "while save_visualizationis True, save_path must be provided."
        if random_color:
            color = np.concatenate([np.random.random(3)], axis=0)
            colors = [color for _ in range(len(masks_list))]
        else:
            cmap = plt.get_cmap("tab10")
            colors = [np.array([*cmap(int(i))[:3]]) for i in obj_id]

        mask_image = [torch.tensor(masks_list[i]).permute(2, 3, 1, 0).numpy() * (
            colors[i].reshape(1, 1, -1)[:, :, :, None]
        ) for i in range(len(masks_list))]

        mask_image = [
            (torch.tensor(mask_image[i]).permute(3, 0, 1, 2) * 255)
            .numpy()
            .astype(np.uint8) for i in range(len(mask_image))
        ]
        mix_image_list = []

        # add mask to video
        width, height = mask_image[0][0].shape[1], mask_image[0][0].shape[0]
        if self.save_visualization:
            result = cv2.VideoWriter(
                save_path, cv2.VideoWriter_fourcc(*"MJPG"), 10, (width, height)
            )

        text_scale = width / 800
        assert video.shape[0] == mask_image[0].shape[0], f"video shape: {video.shape[0]}, mask shape: {mask_image[0].shape[0]}"
        for i in range(video.shape[0]):
            for obj_id in range(len(masks_list)):
                mix_mask = masks_list[obj_id][i][0][:, :, None].repeat(3, axis=2)
                mix_image = np.where(mix_mask, mask_image[obj_id][i], video[i]) if obj_id == 0 else np.where(mix_mask, mask_image[obj_id][i], mix_image)
                # write number on the mask in the image by cv2
                loc = np.where(mix_mask[:,:,0])
                if len(loc[0]) == 0:
                    continue
                loc = (np.mean(loc[0]).astype(int), np.mean(loc[1]).astype(int))

                if loc[0] < 10:
                    loc = (10, loc[1])
                if loc[1] < 10:
                    loc = (loc[0], 10)
                if loc[0] > height - 10:
                    loc = (height - 10, loc[1])
                if loc[1] > width - 10:
                    loc = (loc[0], width - 10)

                cv2.putText(mix_image, str(obj_id+1), (loc[1], loc[0]), cv2.FONT_HERSHEY_TRIPLEX, text_scale, (255, 255, 255), 1, cv2.LINE_AA)

            mix_image_list.append(mix_image)
            if self.save_visualization:
                result.write(mix_image)
                result.release()

        return mix_image_list

    def set_video_list(self, video_list, video_path):
        self.video_list = video_list
        self.inference_state = self.predictor.init_state(video_list, video_path, offload_video_to_cpu=True, offload_state_to_cpu=True)

    def __call__(self, object_points, labels, select_frame, ann_obj_ids):
        masks = []
        # inference_state = self.predictor.init_state(video_path)
        with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
            self.predictor.reset_state(self.inference_state)

            for i in range(len(object_points)):
                self.predictor.add_new_points_or_box(
                    inference_state=self.inference_state,
                    frame_idx=select_frame,
                    obj_id=int(ann_obj_ids[i]),
                    points=object_points[i],
                    labels=labels[i],
                )
                torch.cuda.empty_cache()
            video_segments = (
                {}
            )  # video_segments contains the per-frame segmentation results
            for (
                out_frame_idx,
                out_obj_ids,
                out_mask_logits,
            ) in self.predictor.propagate_in_video(self.inference_state):
                video_segments[out_frame_idx] = {
                    out_obj_id: (out_mask_logits[i] > 0.0)
                    for i, out_obj_id in enumerate(out_obj_ids)
                }
                torch.cuda.empty_cache()
        masks = []
        for i in range(len(object_points)):
            masks.append(
                np.array([
                    video_segments[frame_idx][int(ann_obj_ids[i])].cpu().numpy()
                    for frame_idx in sorted(video_segments.keys())
                ])
            )
        mask = np.stack(masks, axis=0)
        return mask